This repository contains experimental notebooks and scripts created as part of a student research project focused on decoding cognitive states from intracranial EEG (iEEG) data using transformer-based embeddings.
These experiments were developed and tested within the scope of a master's thesis at ITMO University (2025) by G. S. Boykov under the supervision of G. A. Soghoyan.
The goal of this work is to analyze high-dimensional intracranial EEG signals recorded during memory recall tasks using modern dimensionality reduction techniques and latent space embeddings. Transformer-based models are used solely as embedding generators and are not part of this repository. They can be pre-trained to manage the biological individuality of the brain between individuals
*.ipynb: Jupyter notebooks for preprocessing, segmentation, embedding analysis, and visualization..py: Auxiliary scripts to run experiments as a pipelineREADME.md: This file.environment.yml: Optional conda environment (see below).
This repository assumes access to a pretrained embedding model capable of converting preprocessed iEEG segments into sequences of embeddings.
Expected format of embedding machine:
- Input: STFT or similarly preprocessed time-frequency iEEG data
- Output: A tensor of shape (T x D), where T = time bins, D = embedding dimension (e.g., 768)
The embedding model itself is not included in this repository. Any model that satisfies this input-output format can be used.
For embedding generation, any compatible transformer-based model may be used. In our experiments, we tested with models inspired by architectures such as BrainBERT, which can output dense latent representations from time-frequency iEEG inputs.
Please note: The referenced repository is used here solely for academic exploration purposes. No code from the repository is redistributed.
If you use conda, you can recreate the environment using the provided file:
conda env create -f environment.yml
conda activate brainbert